OMOND@DHDEMBL5.BITNET.UUCP (05/15/86)
"Stephan Wasserroth" <unido!vax.hmi.dfn!wasserroth@seismo.CSS.GOV> writes : >On interactive systems like VAX-computers the "response time" is an important >measure for the actual workload of the system. Everybody feels unhappy, if >the response time is high (me too). But it is difficult, to measure the >response time of a CPU. Normally the idle-time is taken as a measure for >the current load on a computer. >Now the question: Do *You* have an idea, how to determine a value describing >the current response time on a computer? Or do *You* have a program for this >purpose? I had the same problem, for which I developed a small program to estimate the load (expressed in %) of total CPU time used by processes at *interactive priority (here 4) and above*. Batch jobs are not counted (avoiding the known LAT problem of an otherwise empty machine appearing to be jam-packed full). I find this gives a good feel for how response will be for online work. The algorithm I use to calculate the load factor uses a smoothed average with probes done once a minute. This is performed by a detached process which takes next to no time itself (on our 8600, cpu time used by the process is 5 seconds in 37 days uptime). The process stores the load factor in the SYS$ANNOUNCE logical name, so that even before you login, you know what to expect. This method also makes it easy to incorporate the load factor into the FINGER program from Richard Garland @ Columbia (thanks Richard, that was a really nice and well-written program !) If anyone is interested in the load estimator, send me an e-mail direct or if enough interest is shown, I can post it to the net (in total it's about 8 blocks). BITNET: Omond@DHDEMBL5.BitNet PSImail: Psi%02624562213002::Omond
zar%Xhmeia.Caltech.Edu@HAMLET.CALTECH.EDU (05/15/86)
[flame if you must] My opinion of a good measure of overall load is to make a daemon perform a simple 30 second (empty VAX) calculation and time the calculation elapsed time over 1 min, 5 min, and 15 min. This would seem to be a better judge of load than indicating # of processes demanding CPU (in COM state). We have lots of batch jobs, and LAVDRIVER is totally ridiculous for determining load (it doesn't start getting sluggish on my VAX until LAV reports numbers above 10). Comments? FTG Zar